No processing of your file contents takes place on any server. You need no data-processing agreement with gottrix to handle confidential or business documents.
Hosted in GermanyGlobal Content Delivery
Origin servers at Hetzner in Germany, delivered worldwide through a content delivery network. Your file contents are never transmitted in the process - they never leave your device.
Shopping lists, meeting notes, a letter from your grandparents or lecture notes: handwriting cannot be searched or copied. This tool reads a photo or scan of German handwriting and turns it into real text. It finds the individual lines itself, recognises each line with a compact neural network and shows the result in a text field where you can fix mistakes directly. Then you copy the text or save it as a TXT file.
The recognition model is called DE-HTR (naeyn/de-htr-web, Apache-2.0 licence). It is about 10 MB, was trained on modern German handwriting and runs via ONNX Runtime as WebAssembly right in your browser. On first use the browser downloads the model together with the runtime once from this website (about 20 MB in total); after that the tool works offline too. Umlauts and ß are part of its character set; there is no dictionary or language model, so mistakes can look like plausible words.
Honest about accuracy: on 601 test lines from the same German handwriting collection the model was trained on, the model author measured a character error rate of about 8.6 percent, roughly one wrong character in twelve. That figure applies to individually cut-out lines. On whole pages, where the tool finds the lines itself, we measured about 21 percent, roughly one wrong character in five. Neat, straight writing on a bright, sharp photo often does better; scrawly writing, slanted lines or shadows do noticeably worse. So always check the text before you use it, especially names, numbers and addresses.
Handwriting is often private: diaries, letters, medical notes. That is why the image stays on your device here. Line finding, recognition and editing run entirely in the browser, nothing is sent to a server, and you need no account.
No file-count or daily limits, no account, no paywall, no watermark. Photos and scans up to 50 megapixels (anything over 4,000 pixels per side is scaled down before recognition) - a browser memory bound; only your device limits the file size (up to 2 GB per run). Many online services limit free use to a few runs per day or require an account; here everything runs locally in your browser - with no daily quota and no sign-up.
Check the recognised text in the field and fix mistakes right there.
Copy the text or download it as TXT.
Limitations: Recognises only present-day GERMAN handwriting; other languages, old scripts such as Kurrent or Sütterlin and printed text are not its target (for print there is "Image to text"). Quality depends heavily on the writing and the photo: straight lines, a single column, a light background and a sharp photo without shadows work best. Slanted or overlapping lines, squared or ruled paper, tables and margin notes are not separated reliably. No dictionary, no formatting: the result is plain text with one line per recognised line and must be checked. On first use the model and its runtime are downloaded once (about 20 MB in total, from this website), then it works offline. Recognition takes roughly half a second to a few seconds per line, depending on the device.
FAQ
Is my image uploaded?
No. Line finding and recognition run entirely in your browser; only the model itself is downloaded from this website the first time. Your image and the recognised text never leave your device.
Which handwriting is recognised?
Modern German handwriting, i.e. cursive and hand-printed letters as people write them today. The model does not recognise Kurrent or Sütterlin, and other languages only by chance, as far as they use the same letters.
How accurate is the recognition?
The model author measured about 8.6 percent wrong characters on 601 test lines from the collection the model was trained on, measured on individually cut-out lines. On whole pages, where the tool finds the lines itself, we measured about 21 percent, roughly one character in five. For your own writing it can be much better or worse; the result is a draft you correct in the text field.
Can I edit the recognised text?
Yes. The result field is editable. Copy and download always take the text as it currently stands in the field, including your corrections.
Does a whole page work?
Yes, but whole pages are recognised noticeably less accurately than single lines: on whole pages we measured about 21 percent wrong characters instead of roughly 6 to 9 percent on single lines. It works best when the lines are reasonably straight and stacked one below the other. The tool separates the lines itself and outputs them in reading order from top to bottom. It does not reliably separate several columns, tables or notes written at an angle.
Is Handwriting to text free - are there any limits?
Yes. You can use Handwriting to text as often as you like: no daily cap, no account, no paywall, no watermark. Fixed bounds: photos and scans up to 50 megapixels (anything over 4,000 pixels per side is scaled down before recognition). All selected files together may be at most 2 GB per run; on devices with little memory this bound is lower, because it follows your device's memory.